Regional climate modeling in the Amazon basin to evaluate fire risk

ABSTRACT Studies regarding deforestation, the hydrological cycle, climate change and fire weather can benefit from the detailed simulations provided by regional climate models (RCM). While much attention has been given to fire activity in the Amazon, few studies have used RCM runs to assess fire risk and variables associated to fire occurrence. We evaluated precipitation, temperature and a fire risk index from the ensemble of Eta model simulations coupled with three different global climate models for the Amazon basin. The RCM runs were compared to reanalysis data for the dry season from 1979 to 2005. The maximum and 2-m temperature fields were underestimated over the entire region, but showed a statistically significant spatial correlation with the reference data. Precipitation was overestimated over the Amazon, in accordance with the major sources of moisture analyzed. The Keetch-Byram drought index (KBDI) was not significantly affected by the bias found in temperature and precipitation, and the ensemble improved relative to the individual member simulations. KBDI estimations performed better with the ensemble of the three evaluated members, however the Eta model showed some limitations. The validation of modeled fire risk could benefit from the use of satellite hotspot data. Furthermore, the KBDI can also be used in the assessment of how climate change interacts with fire activity in the Amazon region.


INTRODUCTION
Global climate models (GCM) are a powerful tool for simulation of present and future climate.In order to function on a global scale, GCMs generally simulate the climate on a horizontal grid of 100-200 km, not allowing to represent some land, land-use, land-ocean and urban features, among others (Ambrizzi et al. 2019).Thus, a GCM has limitations such as the inability to simulate at a high level of detail, the lack of representation of small-scale processes and detailed nearsurface variables (Cabré et al. 2014).One way of surpassing this limitation is to use a regional climate model (RCM) with increased spatial resolution over a smaller domain, forcing it with lateral boundary conditions given by GCMs (Dickinson et al. 1989).Therefore, RCMs can be used as a multipurpose tool for studies that need a higher horizontal resolution (Xuejie et al. 2001;Campbell et al. 2011;Karmalkar et al. 2011;Van Oldenborgh et al. 2013;De Jong et al. 2019).
Due to their resolution, RCMs are expected to be able to represent meso-scale events.The detailed simulation can better represent the spatial and temporal variation of meteorological variables, such as short-duration phenomena, extreme events, small-scale events, cumulus convection, cloud-radiation forcing, and the influence of orography, land-sea and other surface interactions (Wang et al. 2004;Rummukainen et al. 2016).In the Amazon, regional models have already been used to validate the present climate (Chou et al. 2012;Builes-Jaramillo and Pántano 2021), to simulate the effects of deforestation on regional circulation (Ruiz-Vásquez et al. 2020;De Sales et al. 2020) and the hydrological cycle (Gomes et al. 2020), as well as to study climate change (Llopart et al. 2014;Rocha et al. 2015).Detailed simulations of precipitation and temperature can improve the quality of the assessment of changes in climate and the hydrological cycle through representation of changes in regional circulation (Ambrizzi et al. 2019).However, while there are some studies for fire weather and fire risk on a global scale (Liu et al. 2010;Fonseca et al. 2019;Gannon and Steinberg 2021), there is a lack of such studies on a regional scale, including the Amazon region.In the face of climate change scenarios, favorable conditions for fire occurrence may increase in the Amazon basin (Marengo et al. 2018;Vogel et al. 2020).This is particularly important since the Amazon rainforest stores 86 Pg carbon, and almost 80% of this biomass is above ground (Saatchi et al. 2007).When burned, the Amazon rainforest becomes a significant source of carbon for the atmosphere (Nobre and Borma 2009;Balch 2014;Gatti et al. 2021).Therefore, it is essential to understand the relationship between climate conditions and fire in the Amazon.
The vulnerability of the Amazon Forest to fires is enhanced by drought events (Aragão et al. 2014), and this impact has been exacerbated over the years (Anderson et al. 2018).Drought events have a direct impact on carbon emission, which was observed in the 2010, 2015 and 2016 droughts, when anomalous fires in the Amazon were responsible for a combined emission of 0.74 Pg CO 2 (Silva Junior et al. 2019).Moreover, the high spatial variability in precipitation during the drought season affects the distribution of fires in the Amazon basin (Carvalho et al. 2021).As climate change is further affecting precipitation variability and fire occurrence in the Amazon, it becomes increasingly necessary to understand fire dynamics and how it affects the accuracy of simulated fire risk on the regional scale, to understanding the uncertainties that climate models may carry to future climate scenarios.
We aimed to evaluate the physical processes related to optimal meteorological conditions for fire occurrence in the Amazon basin and the applicability of one fire risk index for climate change studies from the regional Eta climate model driven by three global climate models.We estimated the errors and uncertainties of simulations for the present climate during the Amazonian dry period (July, August and September), as well as the model's ability to represent the fire risk index.

Study region
The Amazon basin is the largest hydrographic basin in the world, covering approximately 6.2 million km 2 and encompasses about 40% of the Brazilian territory, as well as parts of Bolivia, Colombia, Ecuador, Guyana, French Guiana, Peru, Suriname and Venezuela.The average annual accumulated rainfall in the region is approximately 2,300 mm year -1 and average temperatures vary between 24 °C and 26 °C, with low thermal amplitude throughout the year (Fisch et al. 1998).One of the defining characteristics of the region is the high spatial and temporal variability of rainfall (Sombroek 2001;Espinoza-Villar et al. 2009), due to the atmospheric systems that act over the region, such as the intertropical convergence zone (Mehta 1998;Wang and Fu 2007), the South Atlantic convergence zone (Kodama 1992), the Bolivian high (Lenters and Cook 1997), the Pacific decadal oscillation and the El Nino-Southern Oscillation (Marengo 2004;Espinoza-Villar et al. 2009).

Regional circulation model
The regional climate model used for this study was the Eta model (Black 1994).One of the main features of the Eta regional model is the Eta vertical coordinate (η), defined by Mesinger (1984), which reduces the error in calculations near steep surfaces of variables such as the strength of the pressure gradient, advective processes and horizontal diffusion (Dereczynski et al. 2000).The Eta model uses Arakawa's E-type horizontal grid (Arakawa and Lamb 1977) and, for this study, it was configured with a horizontal resolution of 20 km and 38 vertical levels.

ACTA AMAZONICA
For the simulations, the Eta model used the parametrization of turbulent diffusion in the planetary boundary layer proposed by Mellor and Yamada (1974).Shortwave radiation (Fels and Schwarzkopf 1975) and longwave radiation (Lacis and Hansen 1974) parametrizations are present.In addition, cumulus parametrization using the Betts-Miller-Janjic scheme (Janjić 1994) and cloud microphysics (Zhao et al. 1997) are used to simulate precipitation.Land processes are represented by the NOAH land surface model (LSM) (Ek et al. 2003).The NOAH model uses four soil layers, 13 vegetation covers and nine different soil types (Hogue et al. 2005).Since the surface layer of the LSM is a combination of the soil and vegetation surface, it is not possible to explicitly calculate some variables such as canopy temperature, carbon fluxes and photosynthetically radiation.However, the LSM has been fitted with various enhancements throughout the years, for canopy conductance soil evaporation, vegetation phenology, surface runoff, infiltration and others (Ek et al. 2003 and references therein).For this simulation, the category of broadleaf-evergreen tree was used to represent the tropical forest, and cultivations to represent deforested areas.
To perform the simulations, the Eta regional model was adapted to use the monthly averages of sea surface temperature (SST) provided by the earth system models (BESM, HadGEM2-ES and MIROC5), to use a 360-day year calendar, in order for the Eta model to be compatible with the HadGEM2-ES lateral boundary conditions.

Earth system models
The Brazilian Earth System Model (BESM) is an earth system model built from a national effort aimed at understanding global climate change, its causes, effects, and impacts on society (Nobre et al. 2013;Capistrano et al. 2020).It was set up at a horizontal resolution of approximately 200 km and 28 vertical levels.In order to represent atmospheric and land surface processes, BESM uses the Brazilian Atmospheric Model (BAM) (Figueroa et al. 2016) and the Simplified Simple Biosphere Model -SSiB (Xue et al. 2001), respectively.The shortwave radiation calculation is based on the CLIRAD-SW-M model (Tarasova et al. 2007), and the Harshvardhan et al. (1987) scheme calculates the longwave radiation.The cloud interaction is based on the scheme described in Slingo (1987), Hou (1990) and Kinter et al. (1997).
The UK Met Office Hadley Centre Global Environmental Model, version 2 (HadGEM2-ES) is an earth system model (Collins et al. 2011).The horizontal resolution of the atmospheric component is N96, approximately 1.875° x 1.25°, with 38 vertical levels.The global dynamic vegetation model TRIFFID (top-down representation of interactive foliage including dynamics) (Cox 2001) was used to describe the terrestrial vegetation and carbon cycle.Oceanic biological and chemical processes were represented by the DiatHadOCC model (Halloran et al. 2010).The UKCA model (United Kingdom Chemistry and Aerosol model) was used to calculate the chemistry of the troposphere.
The Model for Interdisciplinary Research on Climate (MIROC), version 5 is described in detail by Watanabe et al. (2010).The atmospheric spectral component of the model has T85 resolution, which corresponds to approximately 150 km horizontally and has 40 vertical levels.The ocean coupling was performed using the COCO 4.5 model, which has 1° of resolution horizontally and 50 depth levels.The radiative transfers are calculated using the k distribution scheme (Sekiguchi and Nakajima 2008).The model has a cloud microphysics scheme that is coupled with the radiation scheme, and is called SPRINTARS.To represent surface processes MIROC5 uses the MATSIRO scheme (Takata et al. 2003).

Model description and simulation strategy
In this study, the Eta regional climate model, nested with three different earth system models (ESMs), was used to simulate the present climate.The three runs were performed for the period 1960 to 2005, using the initial and boundary conditions of the Brazilian Earth System Model (BESM), the UK Met Office Hadley Centre Global Environmental Model (HadGEM2-ES) and the model for Interdisciplinary Research on Climate (MIROC5).This Eta configuration with Coupled Model Intercomparison Project Phase 5(CMIP5) models was chosen as it has already assessed South America's historical climate (Chou et al. 2014) and also was used in climate change assessments (Brito et al. 2022).The models ran continuously throughout the 45 years, initiating at 00:00 GMT of January 1 st , 1960, in 10-minute time-steps.Precipitation and temperature were given by the model in 6-hour frequency, and later re-calculated to a daily scale.Levels of CO 2 were fixed at 330 ppm.The vegetation map used in the Eta model was obtained from the ProVeg project (Sestini et al. 2002), and updated with deforestation data for the base year of 2015 from Projeto de Monitoramento do Desmatamento da Floresta Amazônica Brasileira por Satélite -PRODES (INPE 2023).

Fire risk index
The Keetch-Byram drought index (KBDI) was used to estimate soil moisture through precipitation and maximum daily temperature.Since soil moisture deficiency and droughts can influence the flammability of vegetation, the KBDI is used as a tool for identifying dry areas that are susceptible to the occurrence of wildfires (Keetch and Byram 1968).The KBDI has been applied in tropical areas (Dolling et al. 2005;Taufik et al. 2015) and specifically in the Amazon biome, with statistically significant results for predicting fire occurrence (Nogueira et al. 2017;Cavalcante et al. 2021).The main advantage of using this index is that it only requires two meteorological variables (daily maximum temperature and ACTA AMAZONICA daily precipitation).It gives values ranging from 0, when there is no soil moisture deficiency, to 800, which denotes absolute soil moisture deficiency (Heim 2002).KBDI values correlate to fire risk in a scale where values from 0 to 200 correspond to low risk, 200 to 400 to moderate risk, 400 to 600 to high risk, and 600 to 800 to very high risk (Liu et al. 2010).
The KBDI is calculated according to equations 1 and 2, with variables expressed following the international unit system (Crane 1982): where dP is the daily precipitation (mm), dQ is the drought factor (mm), KBDI is the moisture deficiency in mm, T is the maximum daily temperature (°C), dt is the time variation, which equals to one day, and R is the average annual precipitation in mm.

Observed data and model validation
The NOAA Climate Prediction Center's (CPC) precipitation and maximum temperature dataset (Xie et al. 2010) was used to assess the ensemble of the Eta simulations with an integration period of 1979-2005, as the CPC dataset begins on January 1979.The CPC dataset is available at a regular spatial resolution of 0.50 x 0.50 degrees.The KBDI, which was calculated with precipitation and maximum temperature simulated by the Eta model, was compared against the KBDI calculated by the same variables from the CPC dataset.Similarly, the simulated near-surface temperature was evaluated using the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 monthly reanalysis data at 0.25 degrees (Hersbach et al. 2020).Maximum temperature was chosen for validation of the regional model because it is one of the most important meteorological variables for fire occurrence.Precipitation was evaluated as it is a natural barrier against fire occurrence and spread.Also, in order to understand how well the model simulates long periods without precipitation and how it impacts the fire risk, we also calculated the number of consecutive dry days (CDD) with a 1mm precipitation threshold.The validation took place over the dry period (July, August and September), using a three-member ensemble (Eta-BESM, Eta-HadGEM2-ES, Eta-MIROC5).For comparison, the bias method was used on mean values and, in order to evaluate the pattern between the reference and simulations, the Taylor diagram (Taylor 2001) was used.

Near-surface temperature
The southern part of the Amazon basin is the area where the models simulated the warmest temperature at 2 m.However, the ensemble models also show a cold bias in the northern and western areas of the basin (Figure 1a).The positive bias was mainly due to the Eta-BESM run (Supplementary Material, Figure S1a), which overestimated the temperature in the entire basin with the exception of the Andes region.On the other hand, Eta-MIROC5 presented the greatest cold bias among the members of the ensemble, although Eta-HadGEM2-ES also demonstrates a cold bias for the northern, western and central regions of the Amazon basin (Supplementary Material, Figures S1b and S1c, respectively).The Taylor diagram shows that all simulations of the Eta model are very close to each other (Figure 1b), indicating that, even when forced with different initial and boundary conditions, all simulations result in a very close temperature pattern output.In addition, the simulations showed high spatial correlation with the reference data, the highest for the Eta-HadGEM2-ES simulation (0.97).The ensemble correlation (0.96) was the

Maximum temperature
The ensemble's cold bias is present throughout the study area (Figure 2a), i.e., the simulations tended to underestimate the maximum temperature in the dry period.The cold bias was lowest in the southeast and northwest of the basin and, consequently, the uncertainty in the simulation is greater there.The Eta model also produced a significant cold bias in the region near the Andes.While the Eta-BESM run (Supplementary Material, Figure S2a) showed a distinct pattern from the Eta-HadGEM2-ES run (Spplementary Material, Figure S2b), which simulated the largest underestimation of maximum temperatures, it had close agreement with the Eta-MIROC5 run (Supplementary Material, Figure S2c).The difference between the standard deviation from the reference data (3.2) and the simulation (approximately 4.8) is a direct consequence of the bias found (Figure 2b).Although the simulations do not show high dispersion among simulated and reference values in the Taylor diagram, the simulated temperature pattern differs from the reference by more than 2 °C.However, as with the near-surface temperature, there was high spatial correlation between reference and simulated maximum temperature, the highest with the three-member ensemble (0.92) and the lowest with Eta-BESM (0.89).

Precipitation
Precipitation was overestimated for this period in almost the entire basin (Figure 3a), with the highest positive bias in the west, over the Andes, with an overestimation of approximately 7 mm day -1 .The overestimation of precipitation is not unique to the Andes, extending to almost the entire area.The Eta-BESM and Eta-HadGEM2-ES simulations showed similar behavior by simulating more than the observed precipitation (Supplementary Material, Figures S3a and S3b), with highest overestimation for Eta-BESM, concentrated in the central and western basin.The Eta-MIROC5, in turn, resulted in more areas with negative (dry) bias (Supplementary Material, Figure S3c) and was the closest to the observed precipitation pattern, as the south of the basin is drier in July-August-September, with higher precipitation in the central and northern regions.

ACTA AMAZONICA
In the Taylor diagram, the precipitation pattern simulated by Eta-BESM deviated more from the observed (reference) values (Figure 3b).Among the four simulations, the ensemble had the highest correlation with the reference values (0.70).

Keetch-Byram drought index
The KBDI showed a well-defined spatial bias distribution (Figure 4b).The ensemble tended to overestimate the KBDI in the central and northern basin, with higher values concentrated in the northeastern sector.In the southeastern basin and in the west, near the Andes, the bias was negative.The simulated KBDI higher than the observed in the northern and central basin despite positive bias for precipitation and negative bias for maximum temperature in this region (Figures 2a and  3a).Even an overestimation of approximately 75 units in the KBDI in the northern region did not change the observed classification of fire risk (Figure 4a).The regions of low and high values for the KBDI calculated from the CPC dataset (Figure 4a) showed positive and negative bias, respectively (Figure 4b).The Eta-BESM run showed predominantly negative bias over the basin (Supplementary Material, Figure S4a, while the Eta-HadGEM2-ES run showed the highest positive bias among the models, and overestimated the KBDI for the entire basin except the Andes (Supplementary Material, Figure S4b).The Eta-MIROC5 run (Supplementary Material, Figure S4c) behaved similarly to the ensemble.
The number of consecutive dry days (CDD) reached more than 40 days of positive bias in the far northeast of the basin, while it was underestimated over the west, near the Andes, and southeast (Figure 5a), where the KBDI was also underestimated.
The Taylor diagram for KBDI showed greatest dispersion among all variables (Figure 5b), reflecting the high variability of simulated KBDI patterns.The Eta-HadGEM2-ES simulation was the closest in standard deviations to the observed data and also showed the highest correlation with the reference values (0.70).The ensemble simulation (approximately 0.69) was very close to the Eta-HadGEM2-ES.

DISCUSSION
The simulations of the present climate have shown significant biases over the Amazon basin.The negative bias in near-surface temperature for the northern basin agrees

ACTA AMAZONICA
with previous simulations using the Eta regional model in the Amazon region.Pesquero et al. (2010) used the Eta model nested within the HadAM3P model and also found underestimation of near-surface temperatures in June, July and August from 1961 to 1970.These coincident patterns could indicate that the underestimation of near-surface temperatures is generated by the Eta model irrespective of its boundary conditions.Furthermore, the prominent negative bias over the Andes can be related to a model response to the topographic configuration and the simulation of high orographic precipitation.The presence of regions where the bias value showed values of zero or very close to zero, however, show that the three-member ensemble is able to improve the results of simulations using individual variables.Despite the uncertainties in the simulations, the Taylor diagrams indicated that the Eta model is able to satisfactorily represent the near-surface temperature in the dry period in the Amazon basin.The underestimation of maximum temperatures over the entire basin agrees with Chou et al. (2014), who found a negative bias for the northern region in both the austral winter and summer, highlighting the tendency of the Eta model to underestimate maximum temperatures for the northern Brazilian Amazon.
While there was an overestimation of precipitation for the entire basin, the bias was greater over the Andes.The presence of the Andes is one of the main driving factors of precipitation in the central and northwestern Amazon basin, as they cause upwards movement of moist air brought by the trade winds (Nobre et al. 1991).Our study confirms that the Eta simulations systematically generate errors in the Andes region, as already shown by Chou et al. (2012 and2014), likely due to the lack of observational data for this region.
In the Amazon, around 32% of the precipitation originates from the basin itself (Staal et al. 2018), therefore the overestimation of local sources of moisture by the Eta model ensemble could lead to increased simulated rainfall.Two moisture sources for precipitation were analyzed to understand this positive bias in precipitation, namely the evapotranspiration and the P-E (precipitation minus evapotranspiration), which indicates the moisture flux convergence (Marengo et al. 2012).We found a positive bias in simulated evapotranspiration and positive values of P-E mean moisture convergence over the region and the P-E was also overestimated over the basin.The behavior of these two moisture sources could explain the bias found in the simulated precipitation, in addition to the errors over the Andes.
The KBDI was overestimated in the central and northern Amazon basin, and underestimated in the Andes and in the southeast.The KBDI considers maximum temperature, precipitation and the accumulation of a deficiency in soil moisture to calculate the fire risk, therefore weather conditions from previous months influence the period evaluated (Liu et al. 2010).Our results suggest that the KBDI does not have a linear relationship with precipitation and maximum temperature.The bias pattern of the KBDI and the CDD were overlayed, as both were overestimated in the northeastern basin, and underestimated in the south.Likewise, in an Eta model forced by HadGEM2-ES for the period 1981-1990, CDD was overestimated in the northeastern region of the Amazon basin, and underestimated in the western and southern regions (Brito et al. 2019).Additionally, using the Eta model at 40 km resolution, Dereczynski et al. (2020) found that CDDs are reduced over the Amazon region.Thus, the nonlinear interactions among precipitation, temperature and CDD likely condition the biases found for the KBDI.
The ensemble simulations seemed to represent KBDI values more accurately at the center, with a poorer representation of the extremes of the variable distribution.The ensemble showed less bias compared to the individual members, suggesting that using the ensemble method can improve the representation of the index and mitigate uncertainties in the simulation of the individual members.
Fires in the Amazon Forest are heavily associated with deforestation and the process of slash-and-burning, when not used properly (Berenguer et al. 2014;Brando et al. 2014).As the KBDI only infers the fire risk through meteorological conditions (Gannon and Steinberg 2021), it may fail to identify areas that are more likely to present fire activity.While forest fragmentation, fuel and ignition are conditions for fire occurrence and spread (Nepstad 1999;Brando et al. 2019), the study of weather conditions that favor to fire activity is equally relevant, not only because fire activity increases in years of extreme drought in the Amazon (Aragão et al. 2014;Anderson et al. 2018;Silva Junior et al. 2019), but also because the Amazon forest is gradually losing its resilience to fire, such as the decrease in cold nights (Balch et al. 2022;Reboita et al. 2022).

CONCLUSIONS
The use of the Eta regional model over the Amazon basin to evaluate fire risk showed significant bias in all variables evaluated.The simulation using the ensemble of near-surface temperature, maximum temperature and precipitation showed improvements when compared to the simulations using individual members and, thus, is suggested as a method to better estimate the fire risk index and mitigate uncertainties in the simulations of the individual members.Our results indicate that the calculation of the KBDI in the Amazon basin is heavily dependent on how well the model simulate precipitation, more specifically the number of consecutive days without rainfall.The simplistic formulation of the KBDI facilitates its calculation and use for fire risk monitoring.In order to reduce the uncertainties of potential fire risk estimation over the region, the Eta model must improve ACTA AMAZONICA the simulation of the climate over the Andes and simulate more accurately the number of consecutive dry days over the entire Amazon region.Information on soil condition and the validation of the KBDI with satellite data on observed fires could increase the accuracy of the fire risk index in the Amazon basin.

Figure 1 .
Figure 1.A -Near-surface temperature bias (°C) from ERA5 versus the ensemble for the dry period (July, August and September) in the Amazon basin from 1979 to 2005.Dotted regions are statistically significant at a 95% confidence level.B -Taylor diagram of the near-surface temperature (°C) for the dry period (July, August and September) in the Amazon basin from 1979 to 2005.The points from 1 to 4 are the different climate simulations and the red star is the reference data.This figure is in color in the electronic version..

Figure 3 .
Figure 3.A -Precipitation bias (mm d -1 ) from CPC versus ensemble, for the dry period (July, August and September) in the Amazon basin from 1979 to 2005.Dotted regions are statistically significant at a 95% confidence level.B -Taylor diagram of the precipitation (mm d -1 ) for the dry period (July, August and September) in the Amazon basin from 1979 to 2005.The points from 1 to 4 are the different climate simulations and the red star is the reference data.This figure is in color in the electronic version.

Figure 2 .
Figure 2. A -Maximum temperature bias (°C) from CPC versus the ensemble, for the dry period (July, August and September) in the Amazon basin from 1979 to 2005.Dotted regions are statistically significant at a 95% confidence level.B -Taylor diagram of the maximum temperature (°C) for the dry period (July, August and September) in the Amazon basin from 1979 to 2005.The points from 1 to 4 are the different climate simulations and the red star is the reference data.This figure is in color in the electronic version.

Figure 5 .
Figure 5.A -Consecutive dry days bias from the CPC versus ensemble, for the dry period in the Amazon basin (July, August and September) from 1979 to 2005.B -Taylor diagram of the Keetch-Byram drought index for the dry period in the Amazon basin (July, August and September) from 1979 to 2005.The points from 1 to 4 are the different climate simulations and the red star is the reference data.This figure is in color in the electronic version..

Figure 4 .
Figure 4. Keetch-Byram drought index calculated from the CPC dataset (A) and calculated from the CPC versus the ensemble data, for the dry period (July, August and September) in the Amazon basin from 1979 to 2005 (B).Dotted regions are statistically significant at a 95% confidence level.This figure is in color in the electronic version..